EADRL: Efficiency-aware adaptive deep reinforcement learning for dynamic task scheduling in edge-cloud environments

In dynamic edge–cloud environments, task scheduling must adapt to fluctuations in workload and resource conditions. This paper presents Efficiency-Aware Adaptive Deep Reinforcement Learning (EADRL), a framework that introduces two key mechanisms: an adaptive learning rate and a dynamic confidence-aware reward adjustment to enable intelligent context-aware scheduling decisions. The learning rate adapts over time to reward trends, improving convergence while maintaining stability under varying load conditions. The reward mechanism aligns with system feedback and incorporates a confidence coefficient derived from recent reward variance to scale Q-value updates. This strategy reduces instability during volatile periods and supports learning under consistent task outcomes. Through these adaptive mechanisms, the EADRL framework fosters resilience and efficiency in task placement across distributed edge and cloud nodes. Experimental evaluations demonstrate that EADRL achieves significant improvements over existing benchmark methods, including DRL-based approaches such as Double DQN (DDQN), Deep Deterministic Policy Gradient (DDPG), and Server Real-Time Performance Deep Reinforcement Learning (SRP-DRL), as well as heuristic-based methods like Best-Fit, Random, and Earliest Idle Time First (EITF). These improvements are evident in key metrics such as task response time, success rate, and load variance. The results validate EADRL's ability to adaptively optimize scheduling policies without prior knowledge of future workloads, making it suitable for scalable and latency sensitive edge-cloud environments. • Proposes EADRL, a deep reinforcement learning framework for efficient and adaptive task scheduling in dynamic edge-cloud environments. • Employs adaptive learning rate tuning and reward optimization to manage real-time workload variation and resource constraints. • Enhances performance with higher task success rate, lower response time, and reduced load variance across edge and cloud tiers. • Outperforms existing methods including DDQN, DDPG, SRP-DRL, Random, EITF, and BEST-FIT in diverse scenarios. • Enables scalable, latency-aware execution for time-critical IoT and edge computing applications.